Ryohei Nakayama PhD, Presenter: Nothing to Disclose

Masaki Ishida MD,PhD, Abstract Co-Author: Nothing to Disclose

Yasutaka Ichikawa MD, Abstract Co-Author: Nothing to Disclose

Mio Uno MD, Abstract Co-Author: Nothing to Disclose

Yoshitaka Goto MD, Abstract Co-Author: Nothing to Disclose

Motonori Nagata MD, PhD, Abstract Co-Author: Nothing to Disclose

Kakuya Kitagawa MD, PhD, Abstract Co-Author: Nothing to Disclose

Hajime Sakuma MD, Abstract Co-Author: Research Grant, Siemens AG Research Grant, Koninklijke Philips NV Research Grant, General Electric Company Research Grant, Bayer AG Research Grant, Guerbet SA

A recent study demonstrated that two-dimensional (2D) learning-based super-resolution (SR) technique can improve image resolution and signal-to-noise ratio (SNR) of whole heart coronary MRA (WHCMRA). However, 2D SR technique cannot increase the through-plane resolution. The purposes of this study were to develop a three-dimensional (3D) SR technique optimized for WHCMRA, and to investigate whether the 3D SR approach can provide high resolution images with improved fidelity and SNR as compared with 2D SR technique.

Free-breathing WHCMRA images were obtained in 46 patients with known or suspected coronary artery disease by using a 1.5T MR system and 32-channel coils, with acquisition resolution of 1.2x1.2x1.5mm and reconstruction resolution of 0.6x0.6x0.75mm. A learning-based 3D SR processing consists of two steps including (1) generation of a 3D dictionary describing relationship between low-resolution (LR) patches and high-resolution (HR) patches, and (2) construction of SR WHCMRA images by embedding 3D patches optimally selected from the dictionary. For evaluating the advantages of the 3D SR processing, WHCMRA images with 0.6x0.6x0.75mm resolution were constructed from the down-sampled WHCMRA images (1.2x1.2x1.5mm) by using 3D-SR, 2D-SR and 3D bi-cubic interpolation (3D-BCI).

The root mean square error between 3D SR images generated from down-sampled WHCMRA and original WHCMRA was 2.75, showing a significant improvement when compared with 2D SR technique (3.28, P < .001) and 3D-BCI (3.57, P < .001). The structural similarity index compared to original WHCMRA was also greater with 3D SR technique (0.982) than with 2D SR technique (0.981, P < .001) and 3D BCI (0.980, P < .001). Although 2D SR approach exhibited significantly improved SNR as compared with 3D-BCI (61.7 +/- 10.5 vs. 49.8 +/-15.7, P < .001), the 3D SR approach proved further improvement in SNR (66.7 +/- 11.5, P = .041 compared with 2D SR).

The 3D SR technique developed in this study can provide high-resolution coronary images with improved fidelity and higher SNR when compared with the 2D SR technique and 3D BCI.

Improved spatial resolution and higher SNR achieved by 3D SR may help to improve the detection of coronary artery stenoses with coronary MRA.

Nakayama, R, Ishida, M, Ichikawa, Y, Uno, M, Goto, Y, Nagata, M, Kitagawa, K, Sakuma, H, Three-dimensional Super-resolution Technique Based on Self-similarity: Usefulness in Whole Heart Coronary Magnetic Resonance Angiography.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14018139.html